RLM (Recurrent Language Model) Integration for LCM Compaction

[viaramb]

Summary

This PR introduces the Recurrent Language Model (RLM) integration into the lossless-claw LCM (Lossless Context Management) compaction system. RLM enables pattern-based summarization that identifies recurring themes, topics, and entities across conversation summaries to produce more coherent and information-dense condensed summaries at higher compaction depths.

What is RLM?

RLM (Recurrent Language Model) is a specialized summarization engine designed for multi-level compaction scenarios. Unlike standard summarization that treats each compaction pass in isolation, RLM:

• Detects recurring patterns across multiple summaries (themes, topics, entities, sentiment shifts)
• Identifies cross-cutting concerns that span multiple conversation segments
• Generates pattern-aware summaries that reference detected patterns instead of repeating them
• Falls back gracefully to standard escalation summarization when patterns are insufficient

This is particularly valuable at depth >= 2 where summaries-of-summaries can become repetitive or lose important thematic connections.

Key Features

1. Pattern-Based Summarization

The RLM engine analyzes RlmSummaryEntry objects (structured metadata extracted from summaries) to detect:

• Recurring themes - Topics that appear across multiple summaries
• Key entities - People, places, concepts mentioned repeatedly
• Sentiment progression - How sentiment evolves across conversation segments
• Unresolvedquestions - Open threads that carry forward

2. Depth-Aware Activation

RLM only activates when depth >= rlmMinDepth (default: 2), ensuring:

• Low-depth compaction uses fast standard summarization
• Higher-depth compaction benefits from pattern analysis
• Configurable threshold based on your use case

3. Seamless Fallback

If RLM pattern detection fails or produces low-confidence results, the system automatically falls back to the standard three-level escalation (normal -> aggressive -> deterministic), ensuring robustness.

4. Heuristic + LLM Hybrid

RLM operates in two modes:

• Heuristic mode (default): Pattern detection via token analysis and frequency counting
• LLM-assisted mode: Optional LLM-based pattern extraction when a completion function is provided

Configuration
Options

Add to your plugins.entries.lossless-claw.config or use environment variables:

Option	Environment Variable	Type	Default	Description
rlmEnabled	LCM_RLM_ENABLED	boolean	false	Enable RLM pattern-based summarization
rlmProvider	LCM_RLM_PROVIDER	string	""	Provider for RLM LLM calls (e.g., "openai")
rlmModel	LCM_RLM_MODEL	string	""	Model for RLM analysis (e.g., "gpt-4o-mini")
rlmMinDepth	LCM_RLM_MIN_DEPTH	number	2	Minimum depth before RLM activates
rlmPatternThreshold	LCM_RLM_PATTERN_THRESHOLD	number	0.7	Confidence threshold for pattern acceptance (0.0-1.0)

Example Configuration

{
"plugins": {
"entries": {
"lossless-claw": {
"config": {
"rlmEnabled": true,
"rlmProvider": "openai",
"rlmModel": "gpt-4o-mini",
"rlmMinDepth": 2,
"rlmPatternThreshold": 0.7
}
}
}
}
}

Files Changed

• src/db/config.ts - Added RLM configuration fields with env var support
• src/rlm/types.ts - Type definitions for RLM entries, patterns, and results
• src/rlm/rlm.ts - Core RlmEngine class with pattern detection and summarization
• src/rlm/index.ts - Module exports
• src/compaction.ts - Integration: RLM instantiation in constructor, summarizeWithEscalation uses RLM for depth >= minDepth
• test/rlm-integration.test.ts - Integration tests for CompactionEngine + RLM
• test/rlm-patterns.test.ts - Unit tests for pattern detection logic

Testing Status

All tests pass:

✓ test/rlm-integration.test.ts (11 tests)
✓ test/rlm-patterns.test.ts (7 tests)

Test Files 2 passed (2)
Tests 18 passed